A Novel Self-Learning Framework for Bladder Cancer Grading Using

Histopathological Images

Gabriel Garcı́a1 , Anna Esteve1,2 , Adrián Colomer1 , David Ramos2 and Valery Naranjo1
1 Instituto de Investigación e Innovación en Bioingenierı́a, Universitat Politècnica de València, 46022, Valencia, Spain
2 Hospital Universitario y Politécnico La Fe, Avinguda de Fernando Abril Martorell, 106, 46026, Valencia, Spain.
arXiv:2106.13559v1 [eess.IV] 25 Jun 2021

Abstract
Recently, bladder cancer has been significantly increased in terms of incidence and mortality. Currently, two sub-
types are known based on tumour growth: non-muscle invasive (NMIBC) and muscle-invasive bladder cancer (MIBC).
In this work, we focus on the MIBC subtype because it is of the worst prognosis and can spread to adjacent organs. We
present a self-learning framework to grade bladder cancer from histological images stained via immunohistochemical
techniques. Specifically, we propose a novel Deep Convolutional Embedded Attention Clustering (DCEAC) which
allows classifying histological patches into different severity levels of the disease, according to the patterns established
in the literature. The proposed DCEAC model follows a two-step fully unsupervised learning methodology to discern
between non-tumour, mild and infiltrative patterns from high-resolution samples of 512 × 512 pixels. Our system out-
performs previous clustering-based methods by including a convolutional attention module, which allows refining the
features of the latent space before the classification stage. The proposed network exceeds state-of-the-art approaches
by 2−3% across different metrics, achieving a final average accuracy of 0.9034 in a multi-class scenario. Furthermore,
the reported class activation maps evidence that our model is able to learn by itself the same patterns that clinicians
consider relevant, without incurring prior annotation steps. This fact supposes a breakthrough in muscle-invasive
bladder cancer grading which bridges the gap with respect to train the model on labelled data.
Keywords: Bladder cancer, tumour budding, unsupervised learning, deep clustering, histopathological images,
self-learning, immunohistochemical staining

1. Introduction
Bladder cancer is an uncontrolled proliferation of the
urothelial bladder tumour cells that entails the devel-
opment of the tumour. A significant increase in adult
incidence and mortality has been observed during the
last years regarding this alteration. In particular, re-
cent studies claim that bladder cancer is the second most
common urinary tract cancer and the fifth most incident
among men in developed countries [1, 2].
Nowadays, the diagnostic procedure of bladder can-
cer includes several time-consuming tests. First, urine
cytology is performed to determine the presence of car-
cinogenic cells [3]. Later, a vesico-prostatic and renal
ultrasound is employed to locate the tumour and get an
idea about its kind of growing, which provides relevant
cues to determine the grade and the prognosis of the
Email address: jogarpa7@i3b.upv.es ()
Preprint submitted to Computers in Biology and Medicine
patient. When the tumour can not be located in the pre-
vious stage, an MRI urogram is carried out to analyze
possible local spread [4]. If there is evidence of blad-
der cancer, the urologist usually performs a cystoscopy
based on the transurethral resection (TUR) technique,
which allows extracting a sample of abnormal bladder
tissue to determine the kind of tumour growth. After
the preparation process, the biopsied tissue is usually
stained with hematoxylin and eosin (H & E) to enhance
the histological properties of the tissue. Finally, an addi-
tional staining process can be adopted to highlight spe-
cial structures associated with the problem under study.
Particularly, the immunohistochemical CK AE1/3 tech-
nique was applied on the histological images used in this
work to highlight the carcinogenic cells by providing a
brown hue when the antigen-antibody binding occurs.
Note that two kinds of bladder cancer, non-muscle in-
vasive (NMIBC) and muscle-invasive (MIBC), are dis-
tinguished depending on its level of invasiveness during
June 28, 2021
the tumour growth within the bladder wall. Currently,
75% and 25% of the bladder cancer cases correspond
to NMIBC and MIBC, respectively [2]. In this study,
we focus on the MIBC category since it leads to the
worst prognosis and favours the spread of the tumour to
adjacent organs. According to [5], the muscle-invasive
bladder cancer (MIBC) does not usually present low-
grade cases of malignancy, but just high-grade urothe-
lial carcinomas, which can be categorized as grade 2 or
3 following the classification criteria proposed by the
World Health Organization (WHO) [6]. Jimenez et al.
[7] described three different histological patterns which
keep correlation with the patient outcome. Specifically,
nodular, trabecular and infiltrative patterns can be found
in the histopathological images stained with CK AE1/3,
as observed in Figure 1. The nodular pattern (yellow
box) is defined by well-delineated tumour cell nests
with a circular shape. Otherwise, the trabecular pat-
tern is characterised by the presence of tumour cells ar-
ranged in interconnected bands. Finally, the infiltrative
pattern is composed of tumour cell strands (red box) or a
small set of isolated cells called buds (blue box). The in-
filtrative pattern, a.k.a. tumour budding, represents the
most aggressive scenario and the worst prognosis for the
patient [8–11]. For that reason, we pigeonholed nodular
and trabecular structures in a single specific class (mild
pattern) to grade the MIBC severity according to the
prognosis of the disease. Also, we considered a non-
tumour pattern (pink box) to cover those cases in which
the patient does not present signs of tumour evidence. In
this way, a multi-class scenario is conducted throughout
the paper to grade the bladder cancer into non-tumour
(NT), mild (M) and infiltrative (I) patterns.
1.1. Related work
An accurate bladder cancer diagnosis supposes a very
time-consuming task for expert pathologists, whose
level of reproducibility is low enough to provide signif-
icant differences in the histological-based interpretation
[12, 13]. For that reason, many studies in the state of
the art have proposed artificial-intelligence algorithms
to help pathologists in terms of cost-effectiveness and
subjectivity ratio. Most of them focused on machine-
learning techniques applied on H & E-stained histolog-
ical images for segmentation [14–16] and classification
[12, 13, 17–22] problems.
Regarding the segmentation-based studies, Lucas et
al. [14] used the popular U-net architecture to segment
normal and malignant cases of bladder images. Then,
they used the common VGG16 network as a backbone
to extract histological features from patches of 224×224
2
pixels. The resulted features were fused with other clini-
cal data to address a classification stage via bidirectional
GRU networks. The proposed algorithm reported accu-
racy of 0.67 for the 5-year survival prediction. In [15],
the authors carried out an end-to-end approach to dis-
cern between MIBC and NMIBC categories from H & E
images. First, they performed a segmentation process
to discriminate the tissue from the background of the
image. Patches of 700 × 700 pixels were used to per-
form both manual and automatic feature extraction. The
hand-crafted learning was conducted via contextual fea-
tures such as nuclear size distribution, crack edge, sam-
ple ratio, etc., whereas the data-driven learning was ad-
dressed via VGG16 and VGG19 architectures. During
the classification stage, different machine-learning clas-
sifiers such as support vector machine (SVM), logis-
tic regression (LR) or random forest (RF), among oth-
ers, were used to determine the kind of bladder tissue.
The hand-driven approaches showcased an outperform-
ing with respect to the deep-learning models, achieving
an accuracy of 91 − 96% depending on the classifier.
Respecting the classification-intended studies, most
of them were focused on H & E-stained histological
images, as before. In [17], the researchers proposed a
multi-class scenario to detect the molecular sub-type in
muscle-invasive bladder cancer (MIBC) cases. They ap-
plied the ResNet architecture on patches of 512 × 512
pixels, achieving results for the area under the ROC
curve (AUC) of 0.89 and 0.87 in terms of micro and
macro-average. Otherwise, in [18], the authors made
use of the Xception network as a feature extractor from
H & E-stained patches of 256 × 256 pixels. Then, an
SVM classifier was implemented to discern between
high and low mutational burden reaching values of 0.73
and 0.75 for accuracy and AUC metrics, respectively.
Harmon et al. [19] proposed a classification scenario
to detect lymph node metastases from H & E patches
of 100 × 100 pixels. A combination of the ResNet-101
architecture with AdaBoost classifiers reported an AUC
of 0.678 at the test time. Another interesting study [12]
carried out a classification approach to categorize the
kind of the tissue into six different classes: urothelium,
stroma, damaged, muscle, blood and background. To
this end, the authors combined supervised and unsuper-
vised deep-learning techniques on patches of 128 × 128
pixels stained with H & E. Specifically, they trained
an autoencoder (AE) from the unlabelled images and
used the encoder network to address the classification
throughout the features extracted from the labelled sam-
ples. They reached multi-class results of 0.936, 0.935
and 0.934% for precision, recall and F1-score metrics,
respectively. One of the more important state-of-the-art
 Nodular

Trabecular

Infiltrative

(a) (b) (c) (d) (e)

Figure 1: Whole Slide Image (WSI) from a patient suffering from muscle-invasive bladder cancer (MIBC) in which different growth patterns are
evidenced. (a) Non-tumour pattern. (b) Nodular arrangement (mild pattern). (c) Trabecular arrangement (mild pattern). (d) Tumour cell strands of
an infiltrative pattern. (e) Isolated tumour cells denoting an infiltrative pattern.

studies focused on H & E histological images of bladder
cancer was carried out in [13]. Zhang et al. collected a
large database of WSIs with the aim of discerning be-
tween low and high grades of the disease. In particular,
they used an autoencoder network to identify possible
areas with cancer. Then, the extracted ROIs of dimen-
sions 1024 × 1024 pixels were input to a Convolutional
Neural Network (CNN) to classify them into low and
high classes. Finally, an average accuracy of 94% was
reported by the proposed system, in comparison to the
84, 3% reached by the pathologists. The findings from
this study reveal that there exists a significant subjectiv-
ity level between experts in the diagnosis from histolog-
ical bladder cancer images, as supported in [12].
It should be noted that, besides the histopathologi-
cal samples, other imaging modalities are also consid-
3
ered in the literature for bladder cancer analysis, e.g.
magnetic resonance imaging (MRI) [16], cystoscopy
[21, 22] or computerized tomography (CT) [20]. Partic-
ularly, Dolz et al. [16] applied deep-learning algorithms
to detect bladder walls and tumour regions from MRI
samples. In [21, 22], different deep-learning architec-
tures were implemented to distinguish between healthy
and bladder cancer patients using cystoscopy samples.
Yang et al. [20] outlined a classification between
NMIBC and MIBC categories from CT images. Addi-
tionally, although immunohistochemical techniques are
widely used in the literature for detecting tumour bud-
ding, most of the state-of-the-art works applied them on
colorectal cancer images [23–26]. However, a large gap
in immunohistochemical-based studies is found in the
literature for bladder cancer diagnosis. As far as we
know, only the study carried out in [27] proposed the use
of immunofluorescence-stained samples to quantify the
tumour budding for MIBC prognosis via machine learn-
ing algorithms. Specifically, the authors aimed to es-
tablish a relationship between tumour budding and sur-
vival evaluated in patients with MIBC. To this end, they
carried out learning strategies based on nuclei detection
and segmentation of the tumour in stroma regions to
count the isolated tumour budding cells. The authors
proposed a survival decision function based on random
forest classifiers, reporting a hazard ratio of 5.44.
1.2. Contribution of this work
To the best of the author’s knowledge, no previous
works have been performed to analyse the severity of
bladder cancer using histological images stained with
cytokeratin AE1-AE3 immunohistochemistry. More-
over, all the state-of-the-art studies focused on super-
vised learning methods to find dependencies between
the inputs and the predicted class [12, 13, 18, 19]. Some
of them [12, 13] also considered the use of unsupervised
techniques in early methodological steps to find possi-
ble ROIs with cancer, but they need labelling data to
build the definitive predictive models. Besides, pattern
recognition tasks aimed to grade bladder cancer have
not been addressed in previous studies.
To fill these gaps in the literature, we present in
this paper a self-learning framework for bladder can-
cer growth pattern, which focuses on fully unsupervised
learning strategies applied on CK AE3/1-stained WSIs.
In particular, we propose a deep convolutional embed-
ded attention clustering (DCEAC) which allows boost-
ing the performance of the classification model without
incurring labelled data. In the literature, deep-clustering
algorithms have demonstrated a high rate of perfor-
mance for image classification [28–30], image segmen-
tation [31], speech separation [32, 33] or data analy-
sis [34], among other tasks. Inspired by [28], we pro-
pose a tailored algorithm capable of competing with
the state-of-the-art results achieved by supervised algo-
rithms. As a novelty, we include a convolutional atten-
tion module to refine the features embedded in the la-
tent space. Additionally, we are the first that focus on
the arrangement of the histological structures contained
in the high-resolution patches to classify them into non-
tumour (NT), mild (M) and infiltrative (I) patterns, ac-
cording to the criteria proposed in [7]. We also com-
pute a class activation map (CAM) algorithm [35] to
evidence how the proposed network pays attention to
those specific structures which match with the clinical
patterns associated with the aggressiveness of bladder
cancer.
4
With all of the above, the proposed end-to-end frame-
work supposes a reliable benchmark for making diag-
nostic suggestions without involving the pathologist ex-
perience, which adds significant value to the body of
knowledge. In summary, the main contributions of this
work are listed below:
• For the first time, we make use of CK AE3/1-
stained images to address the automatic diagnosis
of bladder cancer via machine-learning algorithms.
• We resort to advanced unsupervised deep-learning
techniques to address the bladder cancer grading
without the need for prior annotation steps.
• We propose a novel deep-clustering architecture
able to improve the representation space via con-
volutional attention modules, which derives in a
better-unsupervised classification.
• We based on high-resolution histological patches
to learn specific-bladder cancer patterns and strat-
ify the different severity levels of the disease ac-
cording to the literature.
• Heat maps highlighting decisive areas are reported
to incorporate an explainable component for the
network prediction. This fact provides an inter-
pretability perspective that coincides with the clin-
icians’ criteria.
2. Material
A private database composed of 136 whole-slide
images (WSIs) stained via immunohistochemistry CK
AE1/3 technique was used to accomplish this study.
The WSIs, coming from the Hospital Universitario y
Politécnico La Fe (Valencia, Spain), were digitized at
the highest optical magnification (40×) to leverage the
inherent structure of the bladder patterns associated
with each grade of the disease. Worth noting that a high
image resolution is necessary to achieve an accurate di-
agnosis of bladder cancer. This is because the class de-
pendencies are evidenced in the high frequency of the
image, especially the tumour budding details.
In the first step of the database preparation, an ex-
pert from the Pathological Anatomy Department carried
out a manual segmentation to indicate possible areas of
interest. At this point, it is important to highlight that
the segmentation was performed in a very rough man-
ner, as observed in Figure 2, in order to reduce the ex-
pert’s annotation time as much as possible. From here,
a patching algorithm was applied to extract cropped im-
ages with an optimal block size in terms of computa-
tional efficiency and structural content. Specifically, we
extracted patches of dimensions 512 × 512 pixels, ac-
cording to some of the most recent state-of-the-art stud-
ies focused on histopathological images [17, 36, 37].
Then, a preprocessing step was addressed to discard
the useless regions, corresponding to the background of
the WSI, and select those patches containing more than
75% of annotated tissue. After this, a total of 2995 rep-
resentative patches composed the unsupervised frame-
work. For validation purposes, an expert manually la-
belled each patch as non-tumour (NT), mild (M) or in-
filtrative (I) classes, according to the pattern criteria pre-
viously detailed in Section 1. The labelling process re-
sulted in 763 non-tumour, 1470 mild and 762 infiltrative
cases, as reported in Figure 2. It is essential to remark
that we did not have access to the labelled data during
the training phase since we propose a fully unsupervised
strategy to achieve a self-learning of the patterns. La-
bels were just considered at the test time to evaluate the
models’ performance.
3. Methods
Recently, deep-clustering algorithms have risen to the
forefront of unsupervised image-based techniques since
they allow enhancing the feature learning while improv-
ing the clustering performance in a unified framework
[28]. In this paper, we address a fully unsupervised
self-learning strategy to cluster a large collection of un-
labelled images into K = 3 groups corresponding to
different severity levels of MIBC. Specifically, inspired
by [28], we propose a novel deep convolutional embed-
ded attention clustering (DCEAC) in which the features
are updated in an online manner to learn stable repre-
sentations for the clustering stage. Unlike conventional
approaches [30], the proposed DCEAC algorithm opti-
mizes the latent space by preserving the local structure
of data, which helps to stabilize the clustering-learning
process without distorting the embedding properties.
Self-learning methods aim to learn useful representa-
tions by leveraging the domain-specific knowledge from
the unlabelled data to accomplish downstream tasks.
This training procedure is usually faced by solving pre-
text tasks [38], relational reasoning [39] or contrastive
learning [40] approaches. In our bladder cancer sce-
nario, we advocate for a sequential strategy that resorts
to image reconstruction as an unsupervised pretext task.
Specifically, we conduct a two-step learning methodol-
ogy in which, first, a convolutional autoencoder (CAE)
is trained to incorporate information about the domain
5
properties of the histological patches. In a second phase,
a clustering branch is included at the output of the CAE
bottleneck to provide the class information from the em-
bedded features, which are online updated by re-training
the CAE in a combined network. Below, we detail both
learning steps.
3.1. CAE pre-training
Autoencoder (AE) is one of the most common tech-
niques for data representation, whose aim is to mini-
mize the reconstruction error between inputs X and out-
puts R. AE architectures are composed of two training
stages: encoder fφ (·) and decoder gθ (·), where φ and θ
are learnable parameters. The encoder network applies
a non-linear mapping function to extract a feature space
Z from the input samples X, so that f : X → Z. Then,
the decoder structure is intended to reconstruct the input
data from the embedded representations via R = gθ (Z).
The learning procedure is carried out by minimizing a
reconstruction loss function.
Notice that AE architectures are usually defined by
fully connected layers, intended to reduce the dimen-
sionality of the feature space [30, 34], or by convolu-
tional layers acting as a feature extractor from 2D or 3D
input data [28]. Similarly to [28], we adopted a Con-
volutional AutoEncoder (CAE) architecture to address
the reconstruction of the histological patches as a pre-
text task. However, our CAE differs from the current
literature in a specific aspect of the network: the bot-
tleneck. Unlike Guo et al. [28], who combined flatten
operations with fully-connected layers at the middle of
the CAE, we introduced a convolutional attention mod-
ule through a residual connection to enhance the latent
space for the downstream clustering task. As observed
in Figure 3, we utilized an encoder composed of three
stacked convolutional layers with a 3 × 3 receptive field
(blue boxes). At the bottleneck, we defined an atten-
tion block composed of a tailored autoencoder which
allows refining the embedded features in the spatial di-
mension. Specifically, the proposed module combines
1×1 convolutions (green boxes) with a sigmoid function
(purple layer) intended to recalibrate the inputs. The
inclusion of an identity shortcut forces the network to
stabilize the feature space by propagating larger gradi-
ents to previous layers via skip connections. An addi-
tional 1 × 1 convolutional layer was included to mod-
ify the filter channel without affecting the dimension of
the feature maps. In the decoder stage, we applied reg-
ularization operations between the transpose convolu-
tional layers (yellow-contour boxes) throughout Batch
Normalization to avoid the internal covariate shift [41].
A remarkable factor is that no pooling or upsampling
 CK AE1/3-stained WSIs Patching algorithm Useful patches selection Labelling

Healthy

Grade II

<
512
Grade III

512

Figure 2: Database preparation process. First, a patching algorithm was applied on 136 CK AE1/3-stained WSIs to extract sub-images of 512 × 512
pixels. For validation purposes, the resulting 2995 patches were labelled by an expert as non-tumour (NT), mild (M) or infiltrative (I) pattern to
give rise to a multi-class scenario for bladder cancer grading.

layers were used to adapt the dimensions of the feature
maps after each convolutional step. Instead, we worked
with a stride > 1 both in the encoder and decoder struc-
tures to provide a network with a higher capability of
transformation by learning spatial subsampling.
As observed in Figure 3, given an input set of patches
X = {x1 , x2 , ..., xi , ..., xN }, with N the number of sam-
ples per batch, the encoder network maps each input
xi ∈ R M×M×3 into an embedded feature space zi = fφ (xi )
resulting from the attention module. At the end of the
autoencoder network, the decoder function was trained
to provide a reconstruction map ri = gθ (zi ) trying to
minimize the mean squared error (MSE) between the
input xi and the output ri , according to Equation 1.
Note that the histological patches were resized from
M0 = 512 to M = 128 to alleviate the GPU constraints
during the training of the model.
N rithm in which the decoder structure was discarded dur-
1X
Lr = ||xi − gθ ( fφ (xi ))||2
N i=1
Learning details for the CAE pre-training.
Let X = {X1 , ..., Xb , ..., XB } be the training set com-
posed of 2995 histological patches, the proposed CAE
was trained during  = 200 epochs by applying a learn-
ing rate of 0.5 on B = 94 batches, being Xb ⊂ X a
single batch composed of N = 32 samples. Adadelta
optimizer [42] was used to update the reconstruction
weights trying to minimize the MSE loss function Lr
after each epoch e, as detailed in Algorithm 1.
3.2. DCEAC training
In the pioneer deep-clustering work [30], the authors
proposed a Deep Embedded Clustering (DEC) algo-
Algorithm 1: CAE training.
Data: Unlabelled training data set
X = {X1 , ..., Xb , ..., XB }
Results: Trained convolutional autoencoder
(CAE) parameters φ and θ.
φ, θ ← random;
for e ← 1 to  do
for b ← 1 to B do
X ← Xb ⊂ X;
for i ← 1 to N do
ri ← gθ ( fφ (xi )) ;
PN
Lr ← N1 i=1 ||xi − ri ||2 ;
Update φ, θ using ∇φ,θ Lr
(1) ing the second stage corresponding to the clustering
training. However, Guo et al. [28] demonstrated that
fine-tuning just the encoder network could distort the
feature space and hurt the classification performance.
Instead, they kept the autoencoder untouched under
the statement that AE architectures can avoid embed-
ding distortion by preserving local information of data
[43]. For that reason, we also propose a simultaneous
learning process for both reconstruction and clustering
branches to avoid the corruption of the feature space,
similarly to [28].
Once the CAE was pre-trained in a first stage (Algo-
rithm 1), we incorporated a clustering branch at the out-
put of the CAE bottleneck giving rise to a Deep Convo-
lutional Embedded Attention Clustering (DCEAC) able
6
 𝐸𝑛𝑐𝑜𝑑𝑒𝑟 → 𝑓!(·) 𝐷𝑒𝑐𝑜𝑑𝑒𝑟 → 𝑔" (·)

128×128×64 128×128×64
Stride = 1 Stride = 2 3
3 8×
8 × 12
12 64×64×128 64×64×128 8×
8× Stride = 2 12
12 32×32×256
Stride = 2
Stride = 2

Bottleneck BN BN

𝑥! 𝑟!
32×32×256
Identity shortcut

32×32×64 32×32×1
!!

32×32×32 32×32×256

Figure 3: Architecture of the proposed CAE used for image reconstruction as a pretext task during the learning process.

to provide a soft label with class dependency. From the

q2i, j / i qi, j
P
embedded representations zi = {zi,1 , ..., zi,k , ..., zi,C }, be-
pi, j = P 2 P (4)
ing C = 256 the number of feature maps zi,k ∈ RH×W , j qi, j / i qi, j
we performed a spatial squeeze to obtain a feature vec-
tor z0i ∈ RC which leads to a better label assignment. The learning framework for the proposed DCEAC
As depicted in Figure 4, a Global Average Pooling (Algorithm 2) was conducted by minimizing a custom
(GAP) layer (fadded green) was used as the projection loss function (Equation 5), where Lr and Lc are the re-
function to reduce the feature maps zi,k ∈ RH×W , with construction and clustering losses, respectively. γ > 0
H = W = 32, into the feature vector z0i,k ∈ R1×1 (see is a temperature parameter used to prevent the distortion
Equation 2). of the feature space since γ = 0 would be equivalent to
train just the convolutional autoencoder (CAE) architec-
H W ture, as detailed in Section 3.1.
1 XX
z0i,k = zi,k (h, w) (2)
H × W h=1 w=1 L = Lr + γLc (5)

After the GAP operation, a clustering layer (red box Specifically, the clustering loss was defined as
in Figure 4) was included to map each embedded rep- Kullback-Leibler divergence (KL = (P||Q)) according
resentation z0i into a soft label qi, j , which represents the to Equation 6, whereas the mean squared error (mse)
probability of z0i of belonging to the cluster j. Accord- was used as a reconstruction loss function.
ing to Equation 3, qi, j was calculated via Student’s t- XX pi, j
distribution [44] by keeping the cluster centers {µ j }1K as Lc = pi j log (6)
i j
qi, j
trainable parameters.
As mentioned above, autoencoders are in charge of
(1 + ||z0i − µ j ||2 )−1 preserving the local structure of data, so the clustering
qi, j = P (3)
j (1 + ||zi − µ j || )
0 2 −1 term must provide just a slight contribution to updating
the weights in order to avoid the corruption of the latent
Note that the cluster centres were initialized by run- space. For that reason, we empirically set γ = 0.3 for
ning k-means technique on the embedded features z0i , all the experiments of the training process detailed in
as detailed in Algorithm 2. From here, a normal target Algorithm 2.
distribution pi, j (defined in Equation 4) was used as a
ground truth during the training of the models.
7
 128×128×64
Stride = 1
3 3
8×
12
64×64×128
8× Stride = 2
12 32×32×256
Stride = 2
32×32×256
𝑧!
𝑥!
Identity shortcut
32×32×64 32×32×1
32×32×32 32×32×256
Figure 4: Architecture of the proposed Deep Convolutional Embedded Attention Clustering (DCEAC). The model is trained in an end-to-end
manner by minimizing both reconstruction and clustering loss functions. The reconstruction pretext task stabilizes the feature space zi avoiding the
embedding distortion, whereas the clustering term predicts the soft-class assignments qi .
Learning details for DCEAC training.
As in the previous CAE pre-training, given an input
batch Xb of N = 32 samples, we made use of Adadelta
optimizer with a learning rate of 0.5 to minimize the
custom loss function L. Concerning the software and
hardware aspects, all models were developed using Ten-
sorflow 2.3.1 on Python 3.6. The experiments were per-
formed on a machine with Intel(R) Core(TM) i7-9700
CPU @3.00GHz processor and 16 GB of RAM. For
deep-learning algorithms, a single NVIDIA A100 Ten-
sor Core having cuDNN 7.5 and CUDA Toolkit 10.1
was used.
4. Experimental results
4.1. State of the art
In this section, we show a comparison performance
between the proposed DCEAC model and the most rel-
evant deep clustering-based works of the literature. In
particular, we adapt the study carried out in [30], where
the authors proposed a two-step learning strategy based
on a Deep Embedded Clustering (DEC) model com-
posed of fully connected layers. In the first step, they
trained the autoencoder network to extract knowledge
from the unlabelled images domain. In the second
stage, once the specific-image information was coded,
Xie et al. [30] discarded the decoder structure to di-
rectly address the clustering phase from the learnt fea-
ture space, without considering the reconstruction error.
However, posterior works, such as [28], claimed that
128×128×64
Stride = 2
64×64×128 8×
12
𝑧),+ Stride = 2 8×
12
BN BN
𝑟!
Clustering layer
𝑧′!
𝑞!
𝑧′),+
$ ∑
𝑞!,# / ! 𝑞!,#
$ /∑ 𝑞
∑# 𝑞!,#
𝑝!
! !,#
convolutional autoencoders (CAEs) are more powerful
than fully connected AEs for dealing with images. For
that reason, we adapt the previous DEC methodology by
including convolution operations instead of fully con-
nected layers. To this end, we follow the methodology
exposed in [45], where stacked CAEs were originally
proposed for hierarchical feature extraction. Therefore,
in order to conduct a reliable state-of-the-art compari-
son, we fused both clustering [30] and CAEs architec-
tures [45] to provide a refined DEC model, from now on
called rDEC.
Otherwise, we also replicated the experiments car-
ried out by Guo et al. [28], who proposed a hybrid
learning for deep clustering with convolutional autoen-
coders. The main difference with respect to the previous
rDEC is that [28] keep untouched the decoder term dur-
ing the training of the models giving rise to a hybrid
framework that combines reconstruction Lr and cluster-
ing Lc losses. The idea behind this is that embedded
feature space in rDEC could be distorted by only using
clustering oriented loss. So, they proposed leveraging
the decoder structure to avoid the corruption of the la-
tent space by considering the reconstruction error. Note
that one of the main contributions of Guo et al. [28]
lied in the proposed bottleneck, since they forced the
dimension of the embedded features to be equal to the
number of clusters throughout fully connected layers.
However, this is not scalable to other classification prob-
lems with higher-dimensionality input images or with a
small number of clusters. Specifically, they applied the
8
 Algorithm 2: DCEAC training. ing one of the main own contributions. Hereinafter, we
refer to this approach as rDCEC.
Data: Unlabelled training data set
X = {X1 , ..., Xb , ..., XB } 4.2. Quantitative results
Results: Cluster assignment ŷi for each In this section, we report the unsupervised classi-
histological sample xi . fication performance achieved by the aforementioned
Step 1: Cluster centers intialization rDEC [30] and rDCEC [28] algorithms in comparison
φ, θ ← pre-trained CAE parameters; with our proposed DCEAC model. Also, a conventional
Z ← fφ (X); method based on running k-means algorithm on the fea-
{µ j }Kj=1 ← kmeans(Z); ture space was considered to know how the gap in per-
formance is between the proposed model and traditional
Step 2: DCEAC training
techniques. This conventional approach will be termed
for e ← 1 to  do
as AE+kmeans. As observed in Tables 1 and 2, the
for b ← 1 to B do
comparison is handled by means of different figures of
X ← Xb ⊂ X;
merit, such as sensitivity (SN), specificity (SP), F-score
for i ← 1 to N do
(FS), accuracy (ACC) and area under the ROC curve
zi ← fφ (xi );
(AUC). Particularly, Table 1 shows results per class to
ri ← gθ (zi );
evidence how well the four algorithms classify the 2995
z0i ← GAP(zi );
(1+||z0i −µ j ||2 )−1
histological patches with non-tumour (NT), mild (M)
qi, j ← P 0 2 −1 ; and infiltrative (I) patterns. Besides, Table 2 reports
j (1+||zi −µ j || )
q2i, j / i qi, j the classification results in terms of micro and macro-
P
pi, j ← ;
j qi, j / i qi, j
P 2 P
average. Both metrics provide information about the
PN overall average performance of the classification mod-
Lr ← N1 i=1 ||xi − ri ||2 ;
p els, but micro-average takes into account the unbalanc-
Lc ← i j pi j log qi,i, jj ;
P P
ing between classes, which enables a more faithful per-
L ← Lr + γLc ; spective of the models’ behaviour than macro-average.
Update φ, θ, µ j using ∇φ,θ,µ j L; To enhance the comparison between the learning ap-
Step 3: Label prediction proaches, we represent in Figure 5 the latent space ar-
for b ← 1 to B do ranged by each model with its respective confusion ma-
X ← Xb ⊂ X; trix. Note that, while the confusion matrix gives infor-
for i ← 1 to N do mation about the classification capability of each model,
z0i ← GAP( fφ (xi )); the representation of the embedded space contributes
(1+||z0i −µ j ||2 )−1 to a more comprehensive clustering scenario for the
qi, j ← P 0 2 −1 ; bladder cancer grading. In this way, the T-distributed
j (1+||zi −µ j || )

ŷi ← argmax j (qi, j );
algorithms on the MNIST data set composed of sam-
ples xi ∈ R28×28×1 and provide an embedded space zi
with 10 features, according to the K = 10 number of
clusters. However, in our case, we deal with images
of 128 × 128 × 3 pixels, where the high resolution is
essential for the classification performance, unlike for
MNIST data set. Additionally, we aim to classify the
histological samples into K = 3 classes, so replicating
the architecture of [28] is unfeasible since the decoder
term would be unable of reconstructing the images just
from three feature values. For that reason, to boost a
compelling comparison with [28], we maintain the same
architectures and training details proposed in this work,
but removing the convolutional attention module for be-
9
Stochastic Neighbor Embedding (TSNE) tool was used
to illustrate, in a 2D map, the well and miss-classified
embedded features denoted by spots and crosses, re-
spectively. Green, blue and red colours make reference
to the non-tumour, mild and infiltrative patterns.
4.3. Qualitative results
In an attempt to incorporate an interpretative perspec-
tive for the reported quantitative results, we computed
the class activation maps (CAMs), which allows high-
lighting the regions in which the model pays attention
to predict the class of each sample. This fact usually
can help to find hidden patterns associated with a spe-
cific class or to determine if the label prediction is based
on the same patterns as clinicians. In this way, the re-
ported heatmaps lead to a better understanding of the
embedded feature space by pointing out decisive areas
of the histological patches for the cluster assignment.
 Table 1: Unsupervised classification results per class.
NON-TUMOUR MILD INFILTRATIVE
AE+kmeans rDEC rDCEC DCEAC AE+kmeans rDEC rDCEC DCEAC AE+kmeans rDEC rDCEC DCEAC
SN 0.9345 1 1 0.9987 0.5020 0.8082 0.8952 0.9041 0.5105 0.4659 0.5105 0.6168
SP 0.9870 0.9319 0.9780 0.9978 0.7659 0.8262 0.7862 0.8118 0.6556 0.8782 0.9319 0.9364
FS 0.9475 0.9094 0.9689 0.9961 0.5754 0.8129 0.8458 0.8613 0.4052 0.5112 0.5971 0.6841
ACC 0.9736 0.9492 0.9836 0.9980 0.6364 0.8174 0.8397 0.8571 0.6187 0.7733 0.8247 0.8551

Table 2: Unsupervised classification results in terms of micro and macro-average.

MICRO-AVERAGE MACRO-AVERAGE
AE+kmeans rDEC rDCEC DCEAC AE+kmeans rDEC rDCEC DCEAC
SN 0.6144 0.7699 0.8240 0.8551 0.6490 0.7580 0.8019 0.8399
SP 0.8072 0.8850 0.9120 0.9275 0.8028 0.8788 0.8987 0.9153
FS 0.6144 0.7699 0.8240 0.8551 0.6427 0.7445 0.8039 0.8472
ACC 0.7429 0.8466 0.8827 0.9034 0.7429 0.8466 0.8827 0.9034
AUC 0.7259 0.8184 0.8503 0.8776 0.7259 0.8184 0.8503 0.8776

AE+kmeans rDEC rDCEC DCEAC

NT NT NT NT
True label
True label

True label

True label

M M M M

I I I I

NT M I NT M I NT M I NT M I
Predicted label Predicted label Predicted label Predicted label

Figure 5: Representation of the latent space and confusion matrix derived from the clustering classification reached by each method.

As deduced from Figure 5, the biggest challenge of
the bladder muscle-invasive cancer (MIBC) grading lies
in the distinction of the mild (M) and infiltrative (I) can-
cerous patterns, as expected. For that reason, in Figure
6, we report several examples of heatmaps correspond-
ing to miss-classified samples to elucidate the reason
why the proposed model wrongs. Also, we show exam-
ples of well-predicted CAMs to evidence the relevant
structures in which the network pays attention when
correctly predicting. Specifically, we show five exam-
ples per case to make clear the criteria followed by the
proposed model to determine the class. In the green
frame of Figure 6, we illustrate well-classified mild (a-
e) and infiltrative (f-j) histological patterns. Addition-
ally, in the red frame, we show bladder cancer samples
with a mild pattern miss-classified as tumour budding
(k-o), and vice versa (p-t). The findings from the class
activation maps will be discussed in Section 5.

10
 Well-predicted
mild pattern

(a) (b) (c) (d) (e)

Well-predicted
infiltrative
pattern

(f) (g) (h) (i) (j)

Mild pattern
wrongly predicted
as infiltrative
pattern

(k) (l) (m) (n) (o)

Infiltrative
pattern wrongly
predicted as
mild pattern

(p) (q) (r) (s) (t)

Figure 6: Class activation maps highlighting the regions that the proposed DCEAC model considers relevant for the class prediction. The green
frame refers to well-predicted images with mild (M) and infiltrative(I) patterns, whereas the red frame corresponds to the miss-classified samples
in which the aggressiveness of the disease has been confused.

5. Discussion the outperforming of our model is even more remark-

5.1. About quantitative results
From Table 1, we can observe that all the contrasted
models work well for detecting the non-tumour class,
including the conventional kmeans. However, the pro-
posed DCEAC model reaches the higher performance
for all the metrics, except for sensitivity since the model
miss-classifies one non-tumour sample, as reported in
the DCEAC’s confusion matrix of Figure 5. Regard-
ing the mild class, the proposed model also provides the
best behaviour. Only rDEC surpass a 1% specificity, but
at the cost of compromising a 10% the sensitivity con-
cerning the proposed DCEAC. As observed in Table 1,
11
able when discerning the infiltrative pattern. DCEAC
shows the best results for all the metrics, especially for
sensitivity and F-score exceeding more than a 10% to
the rest of the approaches.
Table 2 reports the overall performance of the mod-
els, in terms of micro and macro-average. As mentioned
above, the micro-average results take into account the
unbalancing between classes, which is an important as-
pect in this study since the samples with mild pattern ap-
pears oversampled. Nevertheless, the proposed DCEAC
model consistently outperforms the rest of the cluster-
ing methods by 2 − 3% across the different both micro
and macro-average metrics, as appreciated in Table 2.
As the final remark concerning the quantitative results,
it should be highlighted that the expert’s decision coin-
cides with the proposed artificial intelligence system in
the 90, 34% of the cases, according to the average accu-
racy.
A reinforcement of the quantitative results is reported
in Figure 5. From the confusion matrices, it is clearly
appreciated by the colour range that all the models tend
to confuse mild and infiltrative cancerous patterns. The
kmeans algorithm presents a very low capability of dis-
cerning between carcinogenic samples since most of the
images are predicted as a mild pattern because of the
oversampling of that class. This changes when deep-
clustering algorithms are outlined. In particular, the
rDEC model improves the classification of carcinogenic
images but compromises a lot the non-tumour class by
miss-classifying samples with an infiltrative pattern. In
this line, the rDCEC model improves the results by no-
tably decreasing the tumour budding samples wrongly
predicted as non-tumour cases. In addition, rDCEC also
increases the number of true positives for tumour sam-
ples. However, this model presents a significant lack
when predicting the infiltrative patterns, since an im-
portant number of them are wrongly labelled as mild.
Unequivocally, the proposed DCEAC model provides
the best classification results. The samples with tumour
budding miss-classified as non-tumour cases decreases
almost to a minimum, unlike in previous cases. More-
over, the number of true positives for mild and infiltra-
tive patterns reaches the highest values, while reducing
the false positives and negatives.
The representation of the embedded feature space of-
fers a visual perspective for the quantitative results. We
can observe that kmeans algorithm is able to roughly
discern between non-tumour and carcinogenic histolog-
ical samples. However, the point cloud is very diffuse
to separate mild and infiltrative classes. Contrarily, the
rDEC model shows a better distribution of the embed-
ded data, although the features relative to each class still
remain close together in the latent space. This improves
in the case of the rDCEC model, where independent
clusters begin to be appreciated. The non-tumour fea-
tures (denoted by the green colour) are unmarked in the
representation space and the embedded tumour samples
begin to scatter towards different cluster classes. In-
disputably, the proposed DCEAC model provides the
best embedding representation since the features are
distributed along the latent space forming independent
clusters according to a specific class. This fact further
strengthens our confidence in the ability of the proposed
model to discern between non-tumour, mild and infiltra-
tive histological patterns.
12
From the previous in-depth analysis of the quantita-
tive and interpretative results, we can clear up several
findings. The first of them is that the use of deep-
learning techniques improves the classification perfor-
mance regarding conventional clustering approaches.
As expected, all the deep clustering-based methods,
i.e. rDEC, DCEC and DCEAC, outperform the base-
line based on the kmeans algorithm. This is because
these models enable a deeper learning stage in which
the embedded features are adjusted to a target distribu-
tion, unlike the kmeans algorithm which modifies the
clusters iteratively without updating the feature learn-
ing. Additionally, we can observe that models with both
reconstruction and clustering branches integrated into a
unified framework provide better results than the rDEC
model, which addresses the learning process in two in-
dependent stages. The reason behind the outperform-
ing of rDCEC and DCEAC with respect to rDEC lies in
the preservation of the local structure of the embedded
data. Since rDCEC and DCEAC models have a con-
nected output between the clustering and reconstruction
stages, the clustering term can transfer class information
to the reconstruction term, which is in charge of updat-
ing the weights of the encoder network. In this way, the
embedded features can be optimized by incorporating
the class prediction without distorting the latent space
thanks to the decoder structure. Finally, the proposed
DCEAC model showcases substantial performance im-
provements regarding the rest of the approaches. This is
due to the inclusion of the convolutional attention block,
which allows refining the latent space to provide more
suitable features for the clustering phase.
5.2. About qualitative results
As appreciated in the class activation maps (CAMs)
reported in Figure 6, the proposed DCEAC model fo-
cuses on tumour cell nests (Figure 6 (a-c)) and tumour
interconnected bands (Figure 6 (d-e)) when predicts
samples with a mild pattern. This fact implies that the
proposed network has learnt by itself to associate nodu-
lar and trabecular structures with a mild pattern of the
disease. Additionally, DCEAC model recognises small
sets of isolated buds (Figure 6 (f-g)) or tumour cell
strands (Figure 6 (h-j)) as characteristic structures of the
infiltrative pattern, a.k.a. tumour budding. These find-
ings are evidenced in the green frame of the heat maps
corresponding to well-predicted samples.
In the case of the wrong predictions (red frame in Fig-
ure 6), we can observe that the proposed network main-
tains consistency when determining the class of each
sample. The histological patches of Figure 6 (k-o) show
an appearance more similar to the infiltrative patterns,
so the network highlights small strands reminiscent of
tumour budding structures. However, the true label as-
signed by the expert for these samples was a mild pat-
tern. At this point, the qualitative results become in-
teresting because the model’ suggestions can lead the
pathologists to reconsider its diagnosis in some doubt-
ful cases, as a second opinion. Besides, the human eye
is susceptible to fatigue, so the proposed system could
help in cases where some patterns have gone unnoticed,
in order to avoid a biased diagnosis.
Otherwise, in the cases of the Figure 6 (p-t), sam-
ples with an infiltrative pattern are wrongly predicted by
the model as mild cases. In these histological patches,
the proposed network focuses on bigger structures re-
lated to nodular or trabecular patterns, but it ignores the
small isolated tumour cells that lead to higher severity
of bladder cancer. From here, we can deduce that, al-
though the final prediction could be wrong, the pattern
recognition accomplished by the model maintains con-
sistency. Notice that the model wrongs because patterns
belonging to a different class coexist in the same histo-
logical patch, so in future research lines, we will face
this long-standing problem.
In summary, the proposed DCEAC model demon-
strates, via class activation maps, a high-confident pre-
diction since it is able to focus on the same patterns
as clinicians, without having previous information from
them. As mentioned above, the opinion from the ex-
pert and the proposed model will match in most of the
cases (in concrete, 90, 34% of the time). Thus, the artifi-
cial intelligence system could help as a computer-aided
system for reviewing processes, which would give rise
to an improvement of the diagnosis quality without in-
volving other experts. Also, the proposed system could
help inexperienced pathologists by suggesting areas of
interest with a convincing label.
6. Conclusion
In this paper, we have proposed a novel self-learning
framework based on deep-clustering techniques to
grade the severity of bladder cancer through histological
samples. Immunohistochemistry staining methods were
applied on the images to enhance the non-tumour, mild
and infiltrative patterns, according to the literature. We
resorted to a hybrid model by combining a clustering
branch along the reconstruction term used as a pretext
task to preserve the local structure of the features. To
stand out from the state of the art, we introduced a con-
volutional attention block that allows refining the fea-
ture space to lead to a better-unsupervised classification.
The proposed Deep Convolutional Embedded Attention
13
Clustering (DCEAC) has demonstrated outperforming
previous clustering-based methods, achieving an aver-
age accuracy of 0.9034 to grade the aggressiveness of
the muscle-invasive bladder cancer (MIBC). Addition-
ally, the reported class activation maps (CAMs) show
that the proposed system is able to learn by itself the
same structures as clinicians to associate the patterns
with the correct severity degree of the disease, without
incurring prior annotation steps. In this line, our fully
unsupervised perspective bridges the gap with respect
to other supervised algorithms, since the proposed sys-
tem does not require expert involvement to be trained.
In future research lines, we will work on improving
the accuracy of tumour samples when structures of dif-
ferent growth patterns appear on the same image. In
addition, we will use more powerful hardware systems
to process entire high-resolution Whole Slide Images to
provide a diagnosis per biopsy, instead of per patch.
Funding
This work has been partially funded by SICAP
project (DPI2016-77869-C2-1-R) and GVA through
project PROMETEO/2019/109. The work of Gabriel
Garcı́a has been supported by the State Research Span-
ish Agency PTA2017-14610-I. The equipment used for
this research has been funded by the European Union
within the operating Program ERDF of the Valen-
cian Community 2014-2020 with the grant number ID-
IFEDER/2020/030.
References
[1] S. Antoni, J. Ferlay, I. Soerjomataram, A. Znaor, A. Jemal,
F. Bray, Bladder cancer incidence and mortality: a global
overview and recent trends, European urology 71 (1) (2017) 96–
108.
[2] L. Lorenzo, Valor pronóstico de la presencia de un componente
tumoral indiferenciado (” tumor budding”) en pacientes con car-
cinoma vesical músculo-invasivo, Ph.D. thesis, Universitat de
València (2018).
[3] G. Feil, A. Stenzl, Pruebas de marcadores tumorales en el cáncer
de vejiga, Actas Urológicas Españolas 30 (1) (2006) 38–45.
[4] S. Sharma, P. Ksheersagar, P. Sharma, Diagnosis and treatment
of bladder cancer, American family physician 80 (7) (2009)
717–723.
[5] A. Stenzl, N. Cowan, M. De Santis, M. Kuczyk, A. Merseburger,
M. Ribal, A. Sherif, J. Witjes, Guı́a clı́nica sobre el cáncer de ve-
jiga con invasión muscular y metastásico, European Association
of Urology (2010).
[6] C. Busch, F. Algaba, The who/isup 1998 and who 1999 systems
for malignancy grading of bladder cancer. scientific foundation
and translation to one another and previous systems, Virchows
Archiv 441 (2) (2002) 105–108.
[7] R. E. Jimenez, E. Gheiler, P. Oskanian, R. Tiguert, W. Sakr,
D. P. Wood Jr, J. E. Pontes, D. J. Grignon, Grading the invasive
 component of urothelial carcinoma of the bladder and its rela-
tionship with progression-free survival, The American journal
of surgical pathology 24 (7) (2000) 980–987.
[8] A. Almangush, M. Karhunen, S. Hautaniemi, T. Salo, I. Leivo,
Prognostic value of tumour budding in oesophageal cancer: a
meta-analysis, Histopathology 68 (2) (2016) 173–182.
[9] E. Karamitopoulou, I. Zlobec, D. Born, A. Kondi-Pafiti, P. Lyk-
oudis, A. Mellou, K. Gennatas, B. Gloor, A. Lugli, Tumour bud-
ding is a strong and independent prognostic factor in pancreatic
cancer, European journal of cancer 49 (5) (2013) 1032–1039.
[10] R. Masuda, H. Kijima, N. Imamura, N. Aruga, Y. Nakamura,
D. Masuda, H. Takeichi, N. Kato, T. Nakagawa, M. Tanaka,
et al., Tumor budding is a significant indicator of a poor prog-
nosis in lung squamous cell carcinoma patients, Molecular
medicine reports 6 (5) (2012) 937–943.
[11] K. Fukumoto, E. Kikuchi, S. Mikami, K. Ogihara, K. Mat-
sumoto, A. Miyajima, M. Oya, Tumor budding, a novel prog-
nostic indicator for predicting stage progression in t1 bladder
cancers, Cancer science 107 (9) (2016) 1338–1344.
[12] R. Wetteland, K. Engan, T. Eftestøl, V. Kvikstad, E. A. Janssen,
Multiclass tissue classification of whole-slide histological im-
ages using convolutional neural networks., ICPRAM 1 (2019)
320–327.
[13] Z. Zhang, P. Chen, M. McGough, F. Xing, C. Wang, M. Bui,
Y. Xie, M. Sapkota, L. Cui, J. Dhillon, et al., Pathologist-level
interpretable whole-slide cancer diagnosis with deep learning,
Nature Machine Intelligence 1 (5) (2019) 236–245.
[14] M. Lucas, I. Jansen, T. G. van Leeuwen, J. R. Oddens, D. M.
de Bruin, H. A. Marquering, Deep learning–based recurrence
prediction in patients with non–muscle-invasive bladder cancer,
European Urology Focus (2020).
[15] P.-N. Yin, K. Kishan, S. Wei, Q. Yu, R. Li, A. R. Haake,
H. Miyamoto, F. Cui, Histopathological distinction of non-
invasive and invasive bladder cancers using machine learning
approaches, BMC medical informatics and decision making
20 (1) (2020) 1–11.
[16] J. Dolz, X. Xu, J. Rony, J. Yuan, Y. Liu, E. Granger,
C. Desrosiers, X. Zhang, I. Ben Ayed, H. Lu, Multiregion seg-
mentation of bladder cancer structures in mri with progressive
dilated convolutional networks, Medical physics 45 (12) (2018)
5482–5493.
[17] A.-C. Woerl, M. Eckstein, J. Geiger, D. C. Wagner, T. Daher,
P. Stenzel, A. Fernandez, A. Hartmann, M. Wand, W. Roth,
et al., Deep learning predicts molecular subtype of muscle-
invasive bladder cancer from conventional histopathological
slides, European urology 78 (2) (2020) 256–264.
[18] H. Xu, S. Park, S. H. Lee, T. H. Hwang, Using transfer learning
on whole slide images to predict tumor mutational burden in
bladder cancer patients, bioRxiv (2019) 554527.
[19] S. A. Harmon, T. H. Sanford, G. T. Brown, C. Yang, S. Mehrali-
vand, J. M. Jacob, V. A. Valera, J. H. Shih, P. K. Agarwal, P. L.
Choyke, et al., Multiresolution application of artificial intelli-
gence in digital pathology for prediction of positive lymph nodes
from primary tumors in bladder cancer, JCO clinical cancer in-
formatics 4 (2020) 367–382.
[20] Y. Yang, X. Zou, Y. Wang, X. Ma, Application of deep learn-
ing as a noninvasive tool to differentiate muscle-invasive blad-
der cancer and non–muscle-invasive bladder cancer with ct, Eu-
ropean Journal of Radiology (2021) 109666.
[21] A. Ikeda, H. Nosato, Y. Kochi, T. Kojima, K. Kawai,
H. Sakanashi, M. Murakawa, H. Nishiyama, Support system of
cystoscopic diagnosis for bladder cancer based on artificial in-
telligence, Journal of endourology 34 (3) (2020) 352–358.
[22] R. Yang, Y. Du, X. Weng, Z. Chen, S. Wang, X. Liu, Automatic
recognition of bladder tumours using deep learning technology
14
and its clinical application, The International Journal of Medical
Robotics and Computer Assisted Surgery (2020) e2194.
[23] F. Prall, H. Nizze, M. Barten, Tumour budding as prognostic
factor in stage i/ii colorectal carcinoma, Histopathology 47 (1)
(2005) 17–24.
[24] A. Lugli, E. Karamitopoulou, I. Panayiotides, P. Karakit-
sos, G. Rallis, G. Peros, G. Iezzi, G. Spagnoli, M. Bihl,
L. Terracciano, et al., Cd8+ lymphocytes/tumour-budding in-
dex: an independent prognostic factor representing a ‘pro-/anti-
tumour’approach to tumour host interaction in colorectal cancer,
British journal of cancer 101 (8) (2009) 1382–1392.
[25] T. Ogawa, T. Yoshida, T. Tsuruta, W. Tokuyama, S. Adachi,
M. Kikuchi, T. Mikami, K. Saigenji, I. Okayasu, Tumor budding
is predictive of lymphatic involvement and lymph node metas-
tases in submucosal invasive colorectal adenocarcinomas and in
non-polypoid compared with polypoid growths, Scandinavian
journal of gastroenterology 44 (5) (2009) 605–614.
[26] I. Zlobec, M. P. Bihl, A. Foerster, A. Rufle, A. Lugli, The impact
of cpg island methylator phenotype and microsatellite instability
on tumour budding in colorectal cancer, Histopathology 61 (5)
(2012) 777–787.
[27] N. Brieu, C. G. Gavriel, I. P. Nearchou, D. J. Harrison,
G. Schmidt, P. D. Caie, Automated tumour budding quantifi-
cation by machine learning augments tnm staging in muscle-
invasive bladder cancer prognosis, Scientific reports 9 (1) (2019)
1–11.
[28] X. Guo, X. Liu, E. Zhu, J. Yin, Deep clustering with convolu-
tional autoencoders, in: International conference on neural in-
formation processing, Springer, 2017, pp. 373–382.
[29] X. Guo, E. Zhu, X. Liu, J. Yin, Deep embedded clustering with
data augmentation, in: Asian conference on machine learning,
PMLR, 2018, pp. 550–565.
[30] J. Xie, R. Girshick, A. Farhadi, Unsupervised deep embedding
for clustering analysis, in: International conference on machine
learning, PMLR, 2016, pp. 478–487.
[31] J. Enguehard, P. O’Halloran, A. Gholipour, Semi-supervised
learning with deep embedded clustering for image classification
and segmentation, IEEE Access 7 (2019) 11093–11104.
[32] J. R. Hershey, Z. Chen, J. Le Roux, S. Watanabe, Deep clus-
tering: Discriminative embeddings for segmentation and sepa-
ration, in: 2016 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), IEEE, 2016, pp. 31–
35.
[33] B. H. Prasetio, H. Tamura, K. Tanno, A deep time-delay em-
bedded algorithm for unsupervised stress speech clustering, in:
2019 IEEE International Conference on Systems, Man and Cy-
bernetics (SMC), IEEE, 2019, pp. 1193–1198.
[34] R. del Amor, A. Colomer, C. Monteagudo, V. Naranjo, A deep
embedded refined clustering approach for breast cancer distinc-
tion based on dna methylation, arXiv preprint arXiv:2102.09563
(2021).
[35] B. Zhou, A. Khosla, A. Lapedriza, et al., Learning deep features
for discriminative localization, in: Proceedings of the IEEE con-
ference on computer vision and pattern recognition, 2016, pp.
2921–2929.
[36] R. del Amor, L. Launet, A. Colomer, A. Moscardó,
A. Mosquera-Zamudio, C. Monteagudo, V. Naranjo, An
attention-based weakly supervised framework for spit-
zoid melanocytic lesion diagnosis in wsi, arXiv preprint
arXiv:2104.09878 (2021).
[37] J. Silva-Rodriguez, A. Colomer, J. Dolz, V. Naranjo, Self-
learning for weakly supervised gleason grading of local patterns,
IEEE journal of biomedical and health informatics (2021).
[38] S. Gidaris, P. Singh, N. Komodakis, Unsupervised represen-
tation learning by predicting image rotations, arXiv preprint
 arXiv:1803.07728 (2018).
[39] M. Patacchiola, A. Storkey, Self-supervised relational reasoning
for representation learning, arXiv preprint arXiv:2006.05849
(2020).
[40] T. Chen, S. Kornblith, M. Norouzi, G. Hinton, A simple frame-
work for contrastive learning of visual representations, in: In-
ternational conference on machine learning, PMLR, 2020, pp.
1597–1607.
[41] S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep
network training by reducing internal covariate shift, in: In-
ternational conference on machine learning, PMLR, 2015, pp.
448–456.
[42] M. D. Zeiler, Adadelta: an adaptive learning rate method, arXiv
preprint arXiv:1212.5701 (2012).
[43] X. Peng, S. Xiao, J. Feng, W.-Y. Yau, Z. Yi, Deep subspace
clustering with sparsity prior., in: IJCAI, 2016, pp. 1925–1931.
[44] L. Van der Maaten, G. Hinton, Visualizing data using t-sne.,
Journal of machine learning research 9 (11) (2008).
[45] J. Masci, U. Meier, D. Cireşan, J. Schmidhuber, Stacked con-
volutional auto-encoders for hierarchical feature extraction, in:
International conference on artificial neural networks, Springer,
2011, pp. 52–59.

15